Google's new AI-powered translation tool

[markcarter]

Announced a couple of days ago, anyone tried this yet?

 

http://www.sciencemag.org/news/2016/09/google-s-new-translation-software-powered-brainlike-artificial-intelligence

[lips]

Where can I try it?

[stapler]

It's live on Google translate now. I entered one of the demonstration sentences from the Google development blog into the current translator tool to check it (https://drive.google.com/file/d/0B4-Ig7UAZe3BSUYweVo3eVhNY3c/view)

 

It's pretty impressive. But it failed on the first sentence I tried (你好土氣), which it translated as "Hello rustic". It seems to be better with more formal/standard Mandarin.

[OneEye]

Yesterday it translated 「你在師大念中文嗎」 as "Do you read Chinese in Normal?" Today, it's translating the same sentence as "Do you read Chinese in Normal University?" So it's improving. 

 

Edit: inspired by stapler's post, I tried 「這個好好吃喔」 and got "This delicious Oh." 

[lips]

I tried it too (Chinese -> English).  For a sentence with a structure similar to English, and containing common words only, it's not bad.  Anything a little bit different, then it's out of whack.  Improvement?  Yes.  More useful?  Not sure.

[Kenny同志]

I am not sure about other language pairs but for Chinese and English, good translation is essentially a rewriting process so Google Translate is unlikely to beat professional E<>C translators unless it could function like the human brain.

[roddy]

There's an interesting (for the particular definitions of interesting we deal with here) graph here showing with actual numbers how much harder the Chinese translation is compared to European pairs. 

 

"Google Translate is unlikely to beat professional E<>C translators unless it could function like the human brain."

First off, that's pretty much *exactly* what they're trying to do with this update. That's why it's "neural". Second, you don't need to be better at translation than someone to steal their work - you just need to be good enough, and cheaper or faster. Eg, Paypal (pdf). 

"Machine translation reduced mechanical work such as typing, terminology lookup and tag placement and allowed post-editors to focus more on style and overall fluency."

Bye bye two translators, hello one editor. 

 

This'll happen later for Chinese, but it will happen. 

[Guest realmayo]

Interesting, I'd think that machines will increasingly improve at breaking down source material but may, as Kenny suggests, continue to struggle with rendering into the target language. So arguably the same as someone translating from their own language into a foreign language that they're not great at.

[Luxi]

Wow! It will save me a lot of time browsing for books.

 

刘慈欣撰写的史诗级科幻巨著！雷军力荐！这是一部典型的硬科幻作品，也是一部伟大的科学著作。他单枪匹马，把中国科幻提升到了世界水平。这是一个关于背叛的故事，也是一个生存与死亡的故事，有时候，比起生存还是死亡来，忠诚与背叛可能更是一个问题。疯狂与偏执，最终将在人类文明的内部异化出怎样的力量？冷酷的星空将如何拷问心中道德？作者试图讲述一部在光年尺度上重新演绎的中国现代史，讲述一个文明二百次毁灭与重生的传奇。 

 

《三体》讲述了地球文明在宇宙中的兴衰历程，书中利用故事模式将社会学、哲学、宗教、人性以及爱情展示得淋漓尽致。在各大书榜中，《三体》系列一直以科幻第一名位居榜首。

 

 

Liu Cixin wrote epic science fiction masterpiece! Lei Jun recommended! This is a typical hard science fiction works, is also a great scientific work. He single-handedly, the Chinese science fiction elevated to the world level. This is a story of betrayal, a story of survival and death, and sometimes, compared to survival or death, loyalty and betrayal may be more of a problem. Crazy and paranoid, and ultimately in the internal alienation of human civilization, what kind of power? Cold Star will be how to torture the hearts of morality? The author tries to recount a Chinese modern history reinterpreted on a light-year scale, telling the story of the destruction and rebirth of a civilization two hundred times.

 

"Three-body" tells the story of the rise and fall of the earth civilization in the course of the universe, the book will use the story mode of sociology, philosophy, religion, humanity and love show most vividly. In the major book list, the "three-body" series has been the first science fiction topped the list.

 

 

This is from a review of Three Bodies It may not be 100% perfect but it is impressive. Further on it translates  三星 as Samsung, probably because it has been corrected to translate as Samsung more often than not. 

[Kenny同志]

This'll happen later for Chinese, but it will happen. 

Machine translation will significantly reduce the demands for human translation, no doubt about that. So I am already thinking about a Plan B.

[Angelina]

The more of the boring, repetitive jobs AI "steals", the better. Allocation of resources might be an issue though, not to mention the danger of turning many people into selfie-taking, cat-videos-sharing "lab rats". 

 

I am pretty interested in AI, despite the fact that I worked hard when I studied conference interpreting. Or is it because of it? AI can be helpful, not only when it comes to translation/interpreting. Why bother with boring jobs? Of course, if anyone's job might be threatened by technology, even people who are not directly involved, should care more. 

 

This is excellent news. Looks like Google used something called Long Short-Term Memory (LSTM). I am only a beginner and learned about something else called backpropagation (BP) today. 

[lips]

It's much improved for getting the gist of a text in a foreign language, as someone mentioned in another thread.  Not that everything will be translated correctly.

 

Using #9 as an example, when the sentence structure was straight forward and parallel to English, it was quite good.  However it also missed some simple things and seemed not to be good in English itself.  As #8 said, "the same as someone translating from their own language into a foreign language that they're not great at."  I'd say that it's not very good in Chinese either.

 

From #9, wonder how could it miss:

 

刘慈欣撰写的史诗级科幻巨著    (missed the 的 and the meaning is different)

也是一部伟大的科学著作    (AI translation should know that adding a conjunction in the English translation makes the English much more natural)

He single-handedly, the Chinese science fiction elevated to the world level    (wrong grammar)

冷酷的星空将如何拷问心中道德  (AI translation should be able to translate this one correctly)

the book will use the story mode of sociology, philosophy, religion, humanity and love show most vividly  (wrong grammar)

In the major book list, the "three-body" series has been the first science fiction topped the list.  (wrong grammar)

 

How could a machine produce English text that is not grammatical?  Didn't it check its own grammar?

 

There are some nicely translated passages though:

 

作者试图讲述一部在光年尺度上重新演绎的中国现代史

The author tries to recount a Chinese modern history reinterpreted on a light-year scale

 

《三体》讲述了地球文明在宇宙中的兴衰历程

"Three-body" tells the story of the rise and fall of the earth civilization in the course of the universe  (But, why was it not Three-body?)

 

Maybe the paradigm is wrong.  #6 was insightful: "good translation is essentially a rewriting process".  Maybe machine translation should be machine reading (and understanding) followed by machine writing, of the whole text.

 

 

 

[li3wei1]

There is no such thing as machine understanding. Not yet, anyway.

A machine can take in the word 'dog' and spit out the equivalents in many languages, or other words for 'dog', or kinds of dogs or parts of dogs. But so far, only a human can see a dog, or hear a dog, or smell a dog, or see a person imitating a dog, and think 'dog'. Or see the word 'doggedly' for the first time, and understand what it means.

[li3wei1]

I'm surprised at the graph that Roddy links to. I would have put the numbers for Chi-Eng in both directions much lower. I've seen stuff that's close to zero, unintelligible rubbish. The only text that Google Translate did well that I've seen was when I needed to find the vision statement in the national biodiversity strategy, so I put a few pages of Chinese text in, and it came out very nearly perfect English. I thought to myself, 'well now I know how they wrote this document'.

[laurenth]

 

Bye bye two translators, hello one editor.

 

As luck would have it, yesterday, I attended a lecture  about post-editing and machine translation. It's pretty obvious that the need for translators is decreasing. Purely human translation is an exception, nowadays. The norm is computer assisted translation. Tomorrow, the norm will be machine translation accompanied (in a dwindling number of cases) by post-editing by a carbon-based life form.

 

I've been a translator for over 20 years and I certainly wouldn't encourage my kids to engage in the same career today. There's a combination of factors at play: people are more and more happy to deal with broken basic English rather than order expensive translations in their mother tongue; CAT has progressed by leaps and bounds; and now machine translation is asymptotically reaching a point where its cheap,  fast and standardised output will be preferred over carefully crafted translations (in fact, it's already the case for the translation of certain technical texts). Machine translation is closer to that tipping point for certain language pairs than others, but it certainly progresses on all fronts.

 

Human translation could become a niche activity, a hobby for craftsmen, like people who bind their own books or brew their own beer. I can contemplate brewing my beer when I'm retired, but it will no longer be possible to pay the rent like that.

 

(Likewise, I'd rather learn Chinese than have a Babel Fish grafted in my brains because (for me) part, if not most, of the interest lies in the human adventure offered by the learning process, rather than in the interesting final by-product of being bilingual.)

[Angelina]

The new Google Translate, at least for the Chinese-English language pair, is less rubbish than the Google Translate you have used. Still rubbish, but less rubbish, more human. Intelligence and understanding are social, you can't separate human understanding from real life. Even when Google Translate can help you translate a text, you should not expect it to be a human being. See it as a tool, like the abacus. 

[roddy]

"I've seen stuff that's close to zero, unintelligible rubbish."

And we'll be seeing it for a long time to come - anything that machine translation hasn't been trained on, it'll likely be rubbish at. And I'm willing to believe computers will never translate a novel. Or at least not a good novel. But for the repetitive technical work which makes up a huge amount of translation work, and where it's worthwhile spending the time to train the machine, it will steal jobs. Maybe not the ad hoc "oh, we need something translated, lets get a freelancer" jobs, but the in-house localisation stuff? That's going. (edit: and key point here - very often the companies control the input - the original is written in-house. They can say to their technical writers "here's the types of structures you can use. Here's the vocab." That's nothing new for a technical writer.")

 

There's very little point putting anything random into Google Translate and saying "look, it doesn't work."* It's like taking a train off the tracks and saying "look,  it doesn't move." Give it two million words of text on the same context and however many linguists and computer scientists (that's your career, kids) it takes, and it'll happily do a decent job.

 

"See it as a tool, like the abacus."

Finally, someone who realises how world-changing it is ;-)

 

*Unless you're a Google engineer, in which case the point of this is "free training".

[Guest realmayo]

I still don't see how, in a Chi-Eng context, this is different from when lots of Chinese people without particularly good English offered to do Chi-Eng work at very cheap rates? Or, did those low rates not stay low for long?

 

See it as a tool, like the abacus.

 

Is the number of accountants in the world higher or lower since the abacus? Since spreadsheets?

 

But seriously: if machine translation gets okayish (but still awkward to read and full of mistakes) then people and organisations who would never have considered having their material translated, because they saw no real need and were unwilling to pay, might now start getting it done by a machine as a matter of course.

 

So, translating content into multiple languages becomes far more common than it is today. Which could then cause increased demand from some of those people to pay for 'good' i.e. edited-by-a-human translations ... which ends up actually meaning more work for people who can translate?

 

These things work in mysterious ways. Who would have thought that after email and the internet killed off the written letter, demand for postal services would rocket (thanks to ebay amazon etc)? Live concerts are way more popular (and lucrative) now than they were before digital music. The value of an original van Gogh painting did not suffer after people could buy cheap poster reproductions, quite the opposite.

 

Okay I have no idea about the world of translation and so on but I'm just saying that in theory there might be room for optimism...

[Angelina]

The value of an original van Gogh painting did not suffer after people could buy cheap poster reproductions, quite the opposite.

 

 

well... 

 

 

there is an essay called The Work of Art in the Age of Mechanical Reproduction, by the author of The Task of the Translator, highly recommended to both Google Translate translators and typewriter+printed dictionary translators (from the good old days of purely human translation )

[imron]

Likewise, I'd rather learn Chinese than have a Babel Fish grafted in my brains because (for me) part, if not most, of the interest lies in the human adventure offered by the learning process

Off-topic: If you haven't already, go and read the short story "Profession" by Isaac Asimov.  You'll enjoy it.